Accurately modeling RNA–ligand interactions is essential for understanding RNA function and for designing RNA-targeted therapeutics. Existing computational approaches mainly focus on predicting either full-length RNA structures or discrete binding sites, lacking explicit modeling of RNA regions that correspond to individual segments within larger, distributed, or structurally ambiguous binding motifs. In this study, we present RLAgent, an interactive agent-based framework that predicts RNA–ligand interactions at the regional level, offering a flexible and detailed modeling perspective, distinct from conventional site-centered approaches. RLAgent defines the RNA–ligand modeling process as an interactive, dialogue-driven workflow. Through a natural language interface, users can specify analysis goals and modeling settings without programming experience. A locally deployed large language model (LLM) serves as the core agent, managing all stages of the modeling pipeline, including data validation, feature representation, model training, performance evaluation, and result visualization. We further evaluated the impact of integrating retrieval-augmented generation (RAG) with RLAgent. The agent-based design of RLAgent reduces technical complexity, improves reproducibility, and makes RNA–ligand modeling more accessible to both computational and experimental researchers.
This study aimed to develop and evaluate an interpretable hierarchical auxiliary classification framework for glaucoma-related fundus image analysis, with emphasis on transparent classification rationale for cases with intermediate aCDR values. A MiT-B2–scSE U-Net was employed to jointly segment the optic disc and optic cup from fundus images, and the aCDR was calculated for initial image-level classification. Samples within predefined low- and high-aCDR ranges were directly classified, whereas samples in the intermediate-aCDR group underwent further analysis using peripapillary vessel segmentation and skeleton-based tortuosity quantification. To improve the reliability of vascular morphology assessment, a distance-aware topology-constrained loss function (DATCL) was introduced in the vessel segmentation stage to improve the segmentation accuracy of peripapillary vessels and improve the reliability of subsequent calculations. On the current REFUGE-based test set, the proposed MiT-B2–scSE U-Net achieved a mean Dice similarity coefficient (Dice) of 0.9254 and a mean Intersection over Union (IoU) of 0.8694 for joint optic cup–disc segmentation. For image-level glaucoma classification, the hierarchical aCDR-plus-vessel strategy achieved an accuracy of 91.25
Medical relation extraction converts unstructured medical texts into structured triples by identifying entities and their semantic relations. However, high annotation costs, limited supervision, and long-tailed relation distributions make zero-shot medical relation extraction particularly challenging, as models need to recognize unseen relation types without relation-specific training examples. Existing methods mainly rely on direct fine-tuning or semantic matching, providing limited supervision for the intermediate decisions that determine the final extraction results. We propose CoT-CC, a Chain-of-Thought reasoning and Contrastive Correction framework for zero-shot medical relation extraction. CoT-CC first distills a four-stage reasoning process consisting of Entity Locator, Entity Classifier, Relation Extractor, and Validator, thereby making entity localization, entity classification, relation inference, and result verification explicit. It then retrieves semantically relevant positive examples and dynamically constructs rule-constrained negative examples targeting two common errors: entity boundary errors and head–tail reversals. By contrasting the initial prediction with these positive and negative references, the correction module refines incorrect entity selections and relation directions before producing the final triples. Experiments on CIC and CMeIE show that CoT-CC achieves better overall performance than representative baselines and shows stable performance across different numbers and partitions of unseen relation types.
Digital burnout is an emerging workforce risk in digitally intensive healthcare systems, yet existing studies remain largely descriptive and provide limited guidance for prioritizing modifiable risk factors. This study develops an explainable artificial intelligence framework for model-based prioritization of digital burnout drivers among healthcare professionals. Cross-sectional survey data were collected from Malaysian healthcare professionals (N = 400). Digital burnout was modelled as a continuous outcome using digital job demands, organizational and personal resources, technostress, and digital minimalism practices as predictors. Linear regression, random forest, and gradient boosting models were evaluated using train-test validation, five-fold cross-validation, and repeated random train-test resampling. SHapley Additive exPlanations were used to characterize predictor contributions and heterogeneity in model predictions. Constrained counterfactual scenario analysis was then conducted to estimate model-based changes in predicted burnout under alternative adjustment scenarios. Gradient boosting demonstrated stable out-of-sample performance, achieving a mean R2 of 0.653 in five-fold cross-validation and 0.660 across 100 repeated 80/20 train-test splits. SHAP analysis identified digital minimalism practices and technostress as the dominant opposing predictors of digital burnout. Single-factor scenario prioritization varied across individuals, risk thresholds, and perturbation magnitudes, whereas the combined DMP-enhancement and technostress-reduction scenario produced the largest model-predicted burnout reduction across all evaluated conditions. The proposed framework advances digital burnout research from prediction toward explainable, resource-aware decision support. The simulated estimates are non-causal and require longitudinal or experimental validation before being interpreted as real-world intervention effects.
Most artificial intelligence models for electrocardiogram (ECG) interpretation rely on task-specific supervised learning, requiring dedicated labeled datasets for each diagnostic task and limiting scalability across diverse cardiac conditions. This study investigated whether vision-language learning from paired ECG recordings and summary reports could enable transferable ECG representations without task-specific supervised training. We developed ECG-VLM, a contrastive vision-language model consisting of a Transformer-based image encoder and a text encoder, trained on 800,034 paired 10-s ECG recordings and summary reports from the MIMIC-IV-ECG database. Cardiac condition detection was performed using natural-language prompts without condition-specific retraining or fine-tuning. Performance was evaluated on an internal test set and externally validated using the LSAS-ECG dataset (45,151 recordings) across 14 cardiac conditions, together with retrieval and subgroup analysis. ECG-VLM achieved diagnostic performance across 14 cardiac conditions, with AUROC ranging from 0.679 to 0.981 on the internal test set and from 0.691 to 0.991 on external validation. Performance was strongest for rhythm abnormalities (mean AUROC 0.948 internally and 0.921 externally) and lower for morphological conditions. Compared with the previously reported signal-based ECG-CLIP framework, ECG-VLM achieved comparable performance for rhythm abnormalities while improving AUPRC for several morphological conditions. Performance remained broadly consistent across external validation and demographic and clinical subgroups. These findings demonstrate the feasibility of vision-language supervision from paired ECG recordings and summary reports for learning transferable ECG representations without task-specific supervised training. This framework provides a basis for future extension of vision-language learning to ECG images encountered in routine clinical practice.
